FEATURE-BASED WATERMARK LOCALIZATION IN DIGITAL CAPTURE SYSTEMS

Abstract

The present disclosures relates generally to digital watermarking and data hiding. One claim recites a method comprising: obtaining data representing captured imagery, the captured imagery depicting packaging including digital watermarking, the digital watermarking including an orientation signal that is detectable in a transform domain; generating a n-dimensional feature set of the data representing captured imagery, the n-dimensional feature set representing the captured imagery in a spatial domain, where n is an integer great than 13; using a trained classifier to predict the presence of the orientation signal in a transform domain from the feature set in the spatial domain. Of course, other claims and combinations are provided too.

Description

RELATED APPLICATIONS

This application claims the benefit of U.S. Provisional Patent Application Nos. 61/856,476, filed Jul. 19, 2013, and 61/918,214, filed Dec. 19, 2013, which are each hereby incorporated herein by reference.

TECHNICAL FIELD

The disclosure relates generally to product packaging, image capture, signal processing, steganographic data hiding and digital watermarking.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows portions of frames captured by a monochromatic camera.

FIG. 2 shows a visual/graphical example of a watermark detection process, including strong detection areas.

FIG. 3 shows a receiver operating characteristic curve of a proposed detector for blocks with watermarked objects.

DETAILED DESCRIPTION

The term “steganography” generally means data hiding. One form of data hiding is digital watermarking. Digital watermarking is a process for modifying media content to embed a machine-readable (or machine-detectable) signal or code into the media content. For the purposes of this application, the data may be modified such that the embedded code or signal is imperceptible or nearly imperceptible to a user, yet may be detected through an automated detection process. Most commonly, digital watermarking is applied to media content such as images, audio signals, and video signals. Watermarking can be incorporated into images or graphics that are then printed, e.g., on product packaging.

Digital watermarking systems may include two primary components: an embedding component that embeds a watermark in media content, and a reading component that detects and reads an embedded watermark (referred to as a “watermark reader,” or “watermark decoder,” or simply as a “reader” or “decoder”). The embedding component (or “embedder” or “encoder”) may embed a watermark by altering data samples representing the media content in the spatial, temporal or some other domain (e.g., Fourier, Discrete Cosine or Wavelet transform domains). The reading component (or “reader” or “decoder”) may analyze target content to detect whether a watermark is present. In applications where the watermark encodes information (e.g., a message or auxiliary payload), the reader may extract this information from a detected watermark.

A watermark embedding process may convert a message, signal or payload into a watermark signal. The embedding process may then combine the watermark signal with media content and possibly other signals (e.g., a transform domain-based orientation pattern or synchronization signal) to create watermarked media content. The process of combining the watermark signal with the media content may be a linear or non-linear function. The watermark signal may be applied by modulating or altering signal samples in a spatial, temporal or transform domain.

A watermark encoder may analyze and selectively adjust media content to give it attributes that correspond to the desired message symbol or symbols to be encoded. There are many signal attributes that may encode a message symbol, such as a positive or negative polarity of signal samples or a set of samples, a given parity (odd or even), a given difference value or polarity of the difference between signal samples (e.g., a difference between selected spatial intensity values or transform coefficients), a given distance value between watermarks, a given phase or phase offset between different watermark components, a modulation of the phase of the host signal, a modulation of frequency coefficients of the host signal, a given frequency pattern, a given quantizer (e.g., in Quantization Index Modulation) etc.

The present assignee's work in steganography, data hiding and digital watermarking is reflected, e.g., in U.S. Pat. Nos. 7,013,021; 6,947,571; 6,912,295; 6,891,959. 6,763,123; 6,718,046; 6,614,914; 6,590,996; 6,408,082; 6,122,403 and 5,862,260, and in published specifications WO 9953428 and WO 0007356 (corresponding to U.S. Pat. Nos. 6,449,377 and 6,345,104). Some 3rd-party work is reflected in, e.g., U.S. Pat. Nos. 7,130,442; 6,208,735; 6,175,627; 5,949,885; 5,859,920. Each of the patent documents identified in this paragraph is hereby incorporated by reference herein in its entirety. Of course, a great many other approaches are familiar to those skilled in the art, e.g., Avcibas, et al., “Steganalysis of Watermarking Techniques Using Images Quality Metrics”, Proceedings of SPIE, January 2001, vol. 4314, pp. 523-531; Dautzenberg, “Watermarking Images,” Department of Microelectronics and Electrical Engineering, Trinity College Dublin, 47 pages, October 1994; Hernandez et al., “Statistical Analysis of Watermarking Schemes for Copyright Protection of Images,” Proceedings of the IEEE, vol. 87, No. 7, July 1999; J. Fridrich and J. Kodovský. Rich models for steganalysis of digital images, IEEE Transactions on Information Forensics and Security, 7(3):868-882, June 2011; J. Kodovský, J. Fridrich, and V. Holub. Ensemble classifiers for steganalysis of digital media, IEEE Transactions on Information Forensics and Security, 7(2):432-444, 2012; and T. Pevný, P. Bas, and J. Fridrich. Steganalysis by subtractive pixel adjacency matrix, IEEE Transactions on Information Forensics and Security, 5(2):215-224, June 2010; I. J. Cox, M. L. Miller, J. A. Bloom, J. Fridrich, and T. Kalker. Digital Watermarking and Steganography, Morgan Kaufman Publishers Inc., San Francisco, Calif., 2007; R. O. Duda, P. E. Hart, and D. H. Stork. Pattern Classification. Wiley Interscience, New York, 2nd edition, 2000; each of which is hereby incorporated herein by reference in its entirety. The artisan is presumed to be familiar with a full range of literature concerning steganography, data hiding and digital watermarking.

Digital watermarking may be used to embed an auxiliary payload into cover media (e.g., images, packaging, graphics, etc.) such that changes to the cover media to convey the digital watermarking remain invisible to humans but allows machines to reliably extract the auxiliary payload even after common signal-processing operations (e.g., noise, filtering, blurring, optical capture). This allows machines to uniquely identify objects they see or hear. Digital watermarking has been used for applications including media content protection, track and trace, etc.

Among other applications, this disclosure addresses an application where digital watermarks are included in consumer packaging (e.g., a soup labels, cereal boxes, etc.). The digital watermark can include information like UPC information, product information, distribution information, retail channel information, and/or an index to such information. Because a large surface area of a package can be watermarked, users no longer need to search for barcode at checkout, thus leading to overall speedup of the checkout process. Assignee's Patent application Ser. Nos. 13/750,752 (published as US 2013-0223673 A1), filed Jan. 25, 2013 and 13/804,413, filed Mar. 14, 2013 (published as US 2014-0112524 A1), and published PCT application No. WO2013033442, which is each hereby incorporated herein by reference in its entirety, discusses related use scenarios. Such retail checkout scenarios are improved when digital watermarking can be located and decoded in a timely manner as watermarked packaging is swiped or moved in front of an optical scanner (or camera). Assuming a fixed or limited time budget allocated for each image frame, limited image frame areas can be examined with a watermark reader to optimize detection time. Due to such constraints, a watermark reader can be configured as a chain of modules, each providing low missed-detection rate information and decreasing false alarms so that un-watermarked image areas can be rejected as early as possible while reducing missing of watermarked area.

One aspect of this disclosure is a pre-watermark detection analysis that quickly analyzes image data and rejects areas that most likely do not contain watermarking. This pre-watermark detection analysis can quickly analyze each area (or subsets of areas) of captured image frames and classify the areas as being watermarked or not. This analysis is preferably less computationally expensive compared to a full watermark read attempt. One technique focuses on image features constructed in the spatial domain to predict whether a transform-domain based digital watermark signal is likely included therein. Preferably, the watermark detector can recognize watermarked areas independent of rotation, scale and pose.

Image areas can be converted into statistical features, e.g., the ‘1a) spam14h,v’ (FIG. 2) in the J. Fridrich et al., “Rich models for steganalysis of digital images,” IEEE Transactions on Information Forensics and Security, 7(3):868-882, June 2011 (“Fridrich”), which is hereby incorporated by reference. A very small portion (including portions of section II, C) from the Fridrich paper is provided below:

C. Co-Occurrence Symmetrization

The individual submodels of the rich image model will be obtained from the 78 co-occurrence matrices computed above by leveraging symmetries of natural images. The symmetries are in fact quite important as they allow us to increase the statistical robustness of the model while decreasing its dimensionality, making it this more compact and improving the performance-to-dimensionality ration. We use the sign-symmetry² as well as the directional symmetry of images. The symmetrization depends on the residual type. All ‘spam’ residuals are symmetrized sequentially by applying the following two rules for all d=(d₁, d₂, d₃, d₄)∈T₄:

C_d←C_d+C_−d,  (5)

C_d← C_d+ C,  (6)

where =(d₄, d₃, d₂, d₁) and −d=(−d₁, −d₂, −d₃, −d₄). After eliminating duplicates from C (which had originally 625 elements), only 169 unique elements remain. ² Sign-symmetry means that taking a negative of an image does not change its statistical properties.

For our purposes, the parameters we chose are quantization step q=3, truncation threshold T=2 and co-occurrence order D=4. An analyzed image frame is first resized, e.g., using bilinear interpolation to 25% of its original size (116 of the original pixels) in order reduce noise, speed up the feature extraction process and to model longer range dependencies. For example, when considering dependencies, image pixel values or residual values are not typically independent, e.g., two neighboring pixels have a likelihood of having the same or nearly the same value. On the other hand, if you take two pixels from opposite side of an image block, the pixels will most likely be very different. Even though the dependencies are weakening with range, there are still some. Being able to utilize (model) these longer range dependencies gives some extra information we can utilize. For example, if you can create a good model of five neighboring pixels it will be better for us than a good model of only two neighboring pixels.

To describe this feature set, we use the symbol X for an M×N grayscale image whose pixel values, X_i,j∈{0, 1, . . . , 255}, are represented using the matrix (x_i,j), i=1, . . . , M, j=1, . . . , N. The horizontal noise residual Z=(z_ij) is computed as z_i,j=x_i,j−x_ij+1. The traditional approach now continues with quantization, rounding and truncation of Z_i,j,

$\begin{array}{cc}{r}_{i,j}={\mathrm{trune}}_T\left(\mathrm{round}\left(\frac{z_{i,j}}{q}\right)\right)& \left(1\right)\end{array}$

and forming a D-dimensional co-occurrence matrix C=(c_{d1, . . . , dD}), d_i∈{−T, . . . , T} from D horizontal neighboring values of r_i,j . The process is then repeated with transpose of the image X to obtain statistics from the vertically neighboring pixels. Both horizontal and vertical co-occurrences are averaged together. By exploiting symmetries in natural images we can reduce the dimensionality by adding up values in co-occurrence bins c_{d1, . . . , dD}, c_{dD, . . . , d1}, c_{−d1, . . . ,−dD}, c_{−dD, . . . , −d1} to obtain the final 169-dimensional feature set.

In order to minimize detection time, e.g., without a conversion of all image data into a transform domain, this feature set characterizes spatial domain image properties. In a case where an embedded watermark signal includes detectable transform domain features (e.g., an orientation signal detectable in the transform domain), we are using the spatial domain feature set to predict the presence of the transform domain orientation signal.

The formulation of the watermark localization problem leads to a binary classification problem to decide whether the feature vector is extracted from an image containing a readable watermark or not. Machine learning is preferably utilized to train a binary classifier on a large amount of labeled training samples so that it can well generalize and successfully classify previously unseen samples. The training establishes a boundary between watermarked and non-watermarked image blocks to enable a pre-watermark analysis module to make fast, accurate decisions.

There are many known classifiers, e.g., Linear Support Vector Machines (L-SVM), Gaussian kernel SVM (G-SVM) and Fisher Linear Discriminant (FLD). Results discussed in this disclosure were obtained using FLD, which exhibited fast classification and simple implementation of Neyman-Pearson Criterion under fixed missed detection rates, but other classifiers could be used in the alternative.

A monochromatic camera with red illumination providing, e.g., 40 1024×1280-pixel frames per second, can be used to obtain image data. Of course other cameras can be used instead. A pre-watermark analysis module obtains image blocks, e.g., 384×384 pixels each, from pre-determined partially overlapping positions (e.g., 2-8 positions).

FIG. 1 shows portions of frames as captured by the monochromatic camera. While the watermark on the tomato soup label (left in FIG. 1) can be seen by some (e.g., grainy texture), it is not possible to read it because of the extreme angle and the shape of the can. The coffee cup (middle in FIG. 1), on the other hand, is not watermarked but the way the air bubbles are arranged may cause a false detection (e.g., False Positive). FIG. 1 (at right image) also shows strong AWGN (additive white Gaussian noise) component cropped from an image with no object present the noise variance is 18.5 and the PSNR is 35.5 dB.

To test the pre-watermark detector module, we acquired around 30,000 images of 20 watermarked test packages (e.g., cans, boxes, bags, etc.) and 10,000 images with non-watermarked objects. In a retail setting, with packages zooming by a checkout scanner or camera, approximately one third of the images captured by the scanner/camera contain pictures of noise and dark background (e.g., no packages present during image capture), so some of these were also included in the training set. Our example training set included approximately 5,000 cover and 5,000 watermarked randomly selected blocks identified by a watermark reader (in this case, a Digimarc Discover reader, provided by Digimarc Corporation, Beaverton, Oreg., USA). Half of the cover blocks were taken from an image database with watermarked objects that were present but cannot be read and half from the image database with the non-watermarked objects.

Most of the cover blocks containing just background were removed for this test.

Two testing sets are used in this example, 10,000 blocks from checkout swipes with watermarked objects and 10,000 blocks from checkout swipes with non-watermarked objects. All images in both sets are preferably completely excluded from the training phase in order to avoid overtraining.

A goal of the disclosed detector is to distinguish between watermarked and non-watermarked blocks for a low fixed missed-detection probability PMD. The results for the thresholds set to achieve PMD 2 {0.01, 0.005, 0.001} are shown in Table 1 and the receiver operating characteristic curve of a proposed detector for block with watermarked objects is plotted in FIG. 3.

A visual (graphical) example of a watermark detection process is shown in FIG. 2 relative to a product packaging. The two right-directional hashed marked boxes highlight the regions where watermarking can be read by a reader. The overlapping hatched areas mark box (middle) is the areas pinpointed by the detector for reading by the watermark reader. The two right most rectangle areas where positively read by a watermark reader, while the left most rectangle area was not.

Our tests shows that the proposed detector is able to reduce the number of images unnecessarily processed by the watermark reader by 60-95% depending on the image content and chosen missed-detection rate. Feature extraction and classification process is efficient to be implemented in software or as part of camera hardware providing the guidance to the detector in real-time.

In some cases a watermark reader outputs a successfully decoded auxiliary payload and a strength indication of a detected synchronization signal (e.g., orientation component). Instead of training a binary classifier that outputs only “watermarked” and “not watermarked,” an alternative classifier is trained with a linear or non-linear regressor that would allow an estimate of detected orientation signal strength directly from the feature set (e.g., from the 169 features). Therefore, it would tell how well an image block is likely to be readable. The watermark reader can prioritize on the blocks with the highest “readability”—strongest estimated orientation strength.

Of course, the invention should not be limited by citing example block sizes, parameters, monochrome scanner, testing sets, etc. For example, a watermark reader can work with much smaller blocks than 384×384 and merge classifier decisions of neighboring blocks to achieve finer watermark localization.

Appendix A of this specification is hereby incorporated by reference. Appendix A provides additional description of, e.g., statistical learning-based methods for localizing watermarked areas corresponding to physical objects depicted in imagery taken by a digital camera (e.g., still or video). In one such example, noise sensitive features and linear classifiers (e.g., a logistic regressor) are used to estimate probabilities of digital watermark presence in some or all subcomponents of an image. These probabilities can then be used to pinpoint areas that are recommended for watermark detection. Preliminary testing shows that this approach is significantly faster than running a watermark detector for each corresponding subcomponent.

CONCLUDING REMARKS

Having described and illustrated the principles of the technology with reference to specific implementations, it will be recognized that the technology can be implemented in many other, different, forms. To provide a comprehensive disclosure without unduly lengthening the specification, applicant hereby incorporates by reference each of the above referenced patent documents in its entirety. Such documents are incorporated in their entireties, even if cited above in connection with specific of their teachings. These documents disclose technologies and teachings that can be incorporated into the arrangements detailed herein (including the arrangements in Appendix A), and into which the technologies and teachings detailed herein can be incorporated.

The methods, processes, components, apparatus and systems described above, and including those in Appendix A, may be implemented in hardware, software or a combination of hardware and software. For example, the watermark encoding processes and embedders may be implemented in software, firmware, hardware, combinations of software, firmware and hardware, a programmable computer, electronic processing circuitry, processors, parallel processors, and/or by executing software or instructions with processor(s) or circuitry. Similarly, watermark data decoding or decoders may be implemented in software, firmware, hardware, combinations of software, firmware and hardware, a programmable computer, electronic processing circuitry, and/or by executing software or instructions with a multi-purpose electronic processor, parallel processors or multi-core processors, and/or other multi-processor configurations.

The methods and processes described above (e.g., watermark detectors) also may be implemented in software programs (e.g., written in C, C++, Visual Basic, Java, Python, Tcl, Perl, Scheme, Ruby, executable binary files, etc.) stored in memory (e.g., a computer readable medium, such as an electronic, optical or magnetic storage device) and executed by an electronic processor (or electronic processing circuitry, hardware, digital circuit, etc.).

The particular combinations of elements and features in the above-detailed embodiments are exemplary only; the interchanging and substitution of these teachings with other teachings in this and the incorporated-by-reference patents are also contemplated.

Claims

1. A method comprising:

obtaining data representing optically captured imagery, the optically captured imagery depicting packaging including digital watermarking, the digital watermarking including an orientation signal that is detectable in a transform domain;

generating an n-dimensional feature set of the data representing the optically captured imagery, the n-dimensional feature set representing the optically captured imagery in a spatial domain, where n is an integer greater than 13;

using a trained classifier to predict the presence of the orientation signal in a transform domain from the feature set in the spatial domain.

2. The method of claim 1 in which said generating comprises calculating difference values between neighboring pixels represented in the data, and quantizing the difference values to yield quantized data.

3. The method of claim 2 in which said generating further comprises rounding the quantized data to the nearest integer to yield rounded data.

4. The method of claim 3 in which said generating further comprises thresholding the rounded data to yield thresholded data.

5. The method of claim 4 in which said generating further comprises accumulating bin-values from the thresholded data.

6. The method of claim 5 in which the n-dimensional feature set comprises a 169-dimensional feature set.

7. The method of claim 1 in which the classifier is trained based on a binary decision of watermarking present or not.

8. The method of claim 1 in which the classifier is trained based on a non-linear regressor corresponding to orientation signal strength.

9. The method of claim 1 in which the classifier is trained based on a linear regressor corresponding to orientation signal strength.

10. An apparatus comprising:

memory for buffering blocks of image data, the image data having been captured with a camera and depicting product packaging;

one or more processors programmed for: generating a spatial domain feature set representation of a block of image data; applying a machine-trained classifier to predict from the spatial domain feature set whether the block of image data includes a transform domain digital watermark signal.

11. The apparatus of claim 10 in which the classifier is trained based on a binary decision of watermarking present or not.

12. The apparatus of claim 10 in which the classifier is trained based on a non-linear regressor corresponding to watermarking signal strength.

13. The apparatus of claim 10 in which the classifier is trained based on a linear regressor corresponding to watermarking signal strength.

14. The apparatus of claim 10 in which said generating comprises:

calculating difference values between neighboring pixels represented in the block of image data, and quantizing the difference values to yield quantized data;

rounding the quantized data to the nearest integer to yield rounded data;

thresholding the rounded data to yield thresholded data; and

accumulating bin-values from the thresholded data.

15. A method comprising:

decomposing data representing an image into subcomponents;

extracting predetermined features from each of the subcomponents;

providing the features to a trained logistic regressor for estimating a probability of a presence of digital watermarking contained in each the subcomponents;

generating a probability map relative to the image based on probabilities generated by the trained logistic regressor; and

selecting n subcomponents from the probability map, where each of the selected n subcomponents has a probability exceeding a predetermined threshold, where n is a positive integer greater than 1.

16. The method of claim 15 further comprising providing only the selected n subcomponents to a digital watermark reader.

17. The method of claim 15 in which the probability map is realized using an electronic table.

18. The method of claim 15 further comprising applying an averaging kernel to the probability map prior to said selecting.

19. A method comprising:

training a binary classifier on a plurality of training samples such that the binary classifier can generalize and classify feature vectors containing digital watermarking or not;

extracting a feature vector from each of a plurality of image areas;

applying the binary classifier to each extracted feature vector to classify the likelihood of a presence of digital watermarking in each image area; and

based on results of said act of applying, directing a digital watermark reader to consider image data from only a subset of the plurality of image areas.